Multimodal Large Language Models (MLLMs) have showcased impressive skills in tasks related to visual understanding and reasoning. Yet, their widespread application faces obstacles due to the high computational demands during both the training and inference phases, restricting their use to a limited audience within the research and user communities. In this paper, we investigate the design aspects of Multimodal Small Language Models (MSLMs) and propose an efficient multimodal assistant named Mipha, which is designed to create synergy among various aspects: visual representation, language models, and optimization strategies. We show that without increasing the volume of training data, our Mipha-3B outperforms the state-of-the-art large MLLMs, especially LLaVA-1.5-13B, on multiple benchmarks. Through detailed discussion, we provide insights and guidelines for developing strong MSLMs that rival the capabilities of MLLMs.

Zhongyi Zhou, Yichen Zhu, Minjie Zhu, Junjie Wen, Ning Liu, Zhiyuan Xu, Weibin Meng, Yaxin Peng, Chaomin Shen, Feifei Feng, Yi Xu. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Continual Learning enables models to learn and adapt to new tasks while retaining prior knowledge. Introducing new tasks, however, can naturally lead to feature entanglement across tasks, limiting the model’s capability to distinguish between new domain data. In this work, we propose a method called Feature Realignment through Experts on hyperSpHere in Continual Learning (Fresh-CL). By leveraging predefined and fixed simplex equiangular tight frame (ETF) classifiers on a hypersphere, our model improves feature separation both intra and inter tasks. However, the projection to a simplex ETF shifts with new tasks, disrupting structured feature representation of previous tasks and degrading performance. Therefore, we propose a dynamic extension of ETF through mixture of experts, enabling adaptive projections onto diverse subspaces to enhance feature representation. Experiments on 11 datasets demonstrate a 2% improvement in accuracy compared to the strongest baseline, particularly in fine-grained datasets, confirming the efficacy of combining ETF and MoE to improve feature distinction in continual learning scenarios.

In this paper, we present DiffusionVLA, a novel framework that integrates autoregressive reasoning with diffusion policies to address the limitations of existing methods: while autoregressive Vision-Language-Action (VLA) models lack precise and robust action generation, diffusion-based policies inherently lack reasoning capabilities. Central to our approach is autoregressive reasoning — a task decomposition and explanation process enabled by a pre-trained VLM — to guide diffusion-based action policies. To tightly couple reasoning with action generation, we introduce a reasoning injection module that directly embeds self-generated reasoning phrases into the policy learning process. The framework is simple, flexible, and efficient, enabling seamless deployment across diverse robotic platforms. We conduct extensive experiments using multiple real robots to validate the effectiveness of DiVLA. Our tests include a challenging factory sorting task, where DiVLA successfully categorizes objects, including those not seen during training. The reasoning injection module enhances interpretability, enabling explicit failure diagnosis by visualizing the model’s decision process. Additionally, we test DiVLA on a zero-shot bin-picking task, achieving \textbf{63.7\% accuracy on 102 previously unseen objects}. Our method demonstrates robustness to visual changes, such as distractors and new backgrounds, and easily adapts to new embodiments. Furthermore, DiVLA can follow novel instructions and retain conversational ability. Notably, DiVLA is data-efficient and fast at inference; our smallest DiVLA-2B runs 82Hz on a single A6000 GPU. Finally, we scale the model from 2B to 72B parameters, showcasing improved generalization capabilities with increased model size.

Diffusion Policy is a powerful technique tool for learning end-to-end visuomotor robot control. It is expected that Diffusion Policy possesses scalability, a key attribute for deep neural networks, typically suggesting that increasing model size would lead to enhanced performance. However, our observations indicate that Diffusion Policy in transformer architecture (DP-T) struggles to scale effectively; even minor additions of layers can deteriorate training outcomes. To address this issue, we introduce Scalable Diffusion Transformer Policy for visuomotor learning. Our proposed method, namely ScaleDP, introduces two modules that improve the training dynamic of Diffusion Policy and allow the network to better handle multimodal action distribution. First, we identify that DPT suffers from large gradient issues, making the optimization of Diffusion Policy unstable. To resolve this issue, we factorize the feature embedding of observation into multiple affine layers, and integrate it into the transformer blocks. Additionally, our utilize non-causal attention which allows the policy network to “see” future actions during prediction, helping to reduce compounding errors. We demonstrate that our proposed method successfully scales the Diffusion Policy from 10 million to 1 billion parameters. This new model, named ScaleDP, can effectively scale up the model size with improved performance and generalization. We benchmark ScaleDP across 50 different tasks from MetaWorld and find that our largest ScaleDP outperforms DP-T with an average improvement of 21.6%. Across 7 real-world robot tasks, our ScaleDP demonstrates an average improvement of 36. 25% over DP-T on four single-arm tasks and 75% on three bimanual tasks. We believe our work paves the way for scaling up models for visuomotor learning. The project page is available at https://scaling-diffusion-policy.github.io/.

Whole-body motion generation from speech audio is crucial for computer graphics and immersive VR/AR. Prior methods struggle to produce natural and diverse whole-body motions from speech. In this paper, we introduce a novel method, named SpeechAct, based on a hybrid point representation and contrastive motion learning to boost realism and diversity in motion generation. Our hybrid point representation leverages the advantages of keypoint representation and surface points of 3D body model, which is easy to learn and helps to achieve smooth and natural motion generation from speech audio. We design a VQ-VAE to learn a motion codebook using our hybrid presentation, and then regress the motion from the input audio using a translation model. To boost diversity in motion generation, we propose a contrastive motion learning method according to the intuitive idea that the generated motion should be different from the motions of other audios and other speakers. We collect negative samples from other audio inputs and other speakers using our translation model. With these negative samples, we pull the current motion away from them using a contrastive loss to produce more distinctive representations. In addition, we compose a face generator to generate deterministic face motion due to the strong connection between the face movements and the speech audio. Experimental results validate the superior performance of our model. The code is available at http://cic.tju.edu.cn/faculty/likun/projects/SpeechAct/index.html.

Neural temporal point processes(TPPs) have shown promise for modeling continuous-time event sequences. However, capturing the interactions between events is challenging yet critical for performing inference tasks like forecasting on event sequence data. Existing TPP models have focused on parameterizing the conditional distribution of future events but struggle to model event interactions. In this paper, we propose a novel approach that leverages Neural Relational Inference (NRI) to learn a relation graph that infers interactions while simultaneously learning the dynamics patterns from observational data. Our approach, the Contrastive Relational Inference-based Hawkes Process (CRIHP), reasons about event interactions under a variational inference framework. It utilizes intensity-based learning to search for prototype paths to contrast relationship constraints. Extensive experiments on three real-world datasets demonstrate the effectiveness of our model in capturing event interactions for event sequence modeling tasks.

Humans interpret scenes by recognizing both the identities and positions of objects in their observations. For a robot to perform tasks such as "pick and place", understanding both what the objects are and where they are located is crucial. While the former has been extensively discussed in the literature that uses the large language model to enrich the text descriptions, the latter remains underexplored. In this work, we introduce the Object-Centric Instruction Augmentation (OCI) framework to augment highly semantic and information-dense language instruction with position cues. We utilize a Multi-modal Large Language Model (MLLM) to weave knowledge of object locations into natural language instruction, thus aiding the policy network in mastering actions for versatile manipulation. Additionally, we present a feature reuse mechanism to integrate the vision-language features from off-the-shelf pre-trained MLLM into policy networks. Through a series of simulated and real-world robotic tasks, we demonstrate that robotic manipulator imitation policies trained with our enhanced instructions outperform those relying solely on traditional language instructions.

The language-conditioned robotic manipulation aims to transfer natural language instructions into executable actions, from simple "pick-and-place" to tasks requiring intent recognition and visual reasoning. Inspired by the dual-process theory in cognitive science—which suggests two parallel systems of fast and slow thinking in human decision-making—we introduce Robotics with Fast and Slow Thinking (RFST), a framework that mimics human cognitive architecture to classify tasks and makes decisions on two systems based on instruction types. Our RFST consists of two key components: 1) an instruction discriminator to determine which system should be activated based on the current user’s instruction, and 2) a slow-thinking system that is comprised of a fine-tuned vision-language model aligned with the policy networks, which allow the robot to recognize user’s intention or perform reasoning tasks. To assess our methodology, we built a dataset featuring real-world trajectories, capturing actions ranging from spontaneous impulses to tasks requiring deliberate contemplation. Our results, both in simulation and real-world scenarios, confirm that our approach adeptly manages intricate tasks that demand intent recognition and reasoning.

We consider a practical application of intelligent vehicles for intersection traffic control. Specially, we study the intersection traffic control problem using reservation-based intersection traffic control system, which utilizes the information exchange between intelligent vehicles and management agents around the intersections to direct traffic, instead of traffic lights. We focus on how to design an effective passing permission (PP) allocation strategy for this system. In this work, with an observation that will cause this system to be inefficient, we propose a novel look-ahead passing permission allocation strategy (LICP) for intersection traffic control. The large-scale testing results show that LICP can make nearly 25% performance improvement on average intersection delay than the previous first come, first serve method (FCFS).